@aws/agentcore 0.21.0 → 0.22.0

package/dist/assets/python/http/strands/base/model/load.py

@@ -1,4 +1,65 @@
 {{#if (eq modelProvider "Bedrock")}}
+{{#if bedrockMantle}}
+import os
+
+from aws_bedrock_token_generator import provide_token
+{{#if (eq mantleApiFormat "chat_completions")}}
+from strands.models.openai import OpenAIModel
+{{else}}
+{{#if mantleProprietary}}
+from strands.models.openai_responses import OpenAIResponsesModel
+{{else}}
+from model.mantle_compat import MantleCompatResponsesModel
+{{/if}}
+{{/if}}
+
+MODEL_ID = "{{modelId}}"
+
+
+def load_model():
+    """
+    Get a Bedrock Mantle model client. These OpenAI-compatible models (e.g. openai.gpt-5.5,
+    openai.gpt-oss-120b) are served via the Bedrock Mantle endpoint, NOT the Converse API — so they
+    are invoked through an OpenAI-style client authenticated with a short-lived Bedrock bearer token.
+    Region is read from AWS_REGION (set by the AgentCore runtime).
+    """
+    region = os.environ.get("AWS_REGION", os.environ.get("AWS_DEFAULT_REGION", "us-east-1"))
+    token = provide_token(region=region)
+    {{#if mantleProprietary}}
+    # Proprietary OpenAI models only work on the /openai/v1 Mantle path.
+    base_url = f"https://bedrock-mantle.{region}.api.aws/openai/v1"
+    {{else}}
+    # Open-source OpenAI models (gpt-oss-*) only work on the /v1 Mantle path.
+    base_url = f"https://bedrock-mantle.{region}.api.aws/v1"
+    {{/if}}
+    client_args = {"api_key": token, "base_url": base_url}
+
+    params = {}
+    {{#if modelMaxTokens}}
+    {{#if (eq mantleApiFormat "chat_completions")}}
+    params["max_completion_tokens"] = {{modelMaxTokens}}
+    {{else}}
+    params["max_output_tokens"] = {{modelMaxTokens}}
+    {{/if}}
+    {{/if}}
+    {{#if modelTemperature}}
+    params["temperature"] = {{modelTemperature}}
+    {{/if}}
+    {{#if modelTopP}}
+    params["top_p"] = {{modelTopP}}
+    {{/if}}
+    {{#if (eq mantleApiFormat "chat_completions")}}
+    return OpenAIModel(client_args=client_args, model_id=MODEL_ID, params=params)
+    {{else}}
+    # Responses API: Mantle does not persist responses, so disable server-side storage.
+    params["store"] = False
+    {{#if mantleProprietary}}
+    return OpenAIResponsesModel(client_args=client_args, model_id=MODEL_ID, params=params)
+    {{else}}
+    return MantleCompatResponsesModel(client_args=client_args, model_id=MODEL_ID, params=params)
+    {{/if}}
+    {{/if}}
+{{else}}
 from strands.models.bedrock import BedrockModel
 
 
@@ -6,6 +67,7 @@ def load_model() -> BedrockModel:
     """Get Bedrock model client using IAM credentials."""
     return BedrockModel(model_id="{{#if modelId}}{{modelId}}{{else}}global.anthropic.claude-sonnet-4-5-20250929-v1:0{{/if}}")
 {{/if}}
+{{/if}}
 {{#if (eq modelProvider "Anthropic")}}
 import os
 
@@ -121,3 +183,57 @@ def load_model() -> GeminiModel:
         model_id="{{#if modelId}}{{modelId}}{{else}}gemini-2.5-flash{{/if}}",
     )
 {{/if}}
+{{#if (eq modelProvider "LiteLLM")}}
+import os
+{{#if litellmAdditionalParams}}
+import json
+{{/if}}
+
+from strands.models.litellm import LiteLLMModel
+{{#if identityProviders.[0].name}}
+from bedrock_agentcore.identity.auth import requires_api_key
+
+IDENTITY_PROVIDER_NAME = "{{identityProviders.[0].name}}"
+IDENTITY_ENV_VAR = "{{identityProviders.[0].envVarName}}"
+
+
+@requires_api_key(provider_name=IDENTITY_PROVIDER_NAME)
+def _agentcore_identity_api_key_provider(api_key: str) -> str:
+    """Fetch API key from AgentCore Identity."""
+    return api_key
+
+
+def _get_api_key() -> str:
+    """
+    Uses AgentCore Identity for API key management in deployed environments.
+    For local development, run via 'agentcore dev' which loads agentcore/.env.
+    """
+    if os.getenv("LOCAL_DEV") == "1":
+        api_key = os.getenv(IDENTITY_ENV_VAR)
+        if not api_key:
+            raise RuntimeError(
+                f"{IDENTITY_ENV_VAR} not found. Add {IDENTITY_ENV_VAR}=your-key to .env.local"
+            )
+        return api_key
+    return _agentcore_identity_api_key_provider()
+{{/if}}
+
+
+
+
+def load_model() -> LiteLLMModel:
+    """Get a LiteLLM model client (proxies to the provider encoded in model_id)."""
+    client_args = {}
+    {{#if identityProviders.[0].name}}
+    client_args["api_key"] = _get_api_key()
+    {{/if}}
+    {{#if litellmApiBase}}
+    client_args["api_base"] = {{safeJson litellmApiBase}}
+    {{/if}}
+    params = {{#if litellmAdditionalParams}}json.loads({{pyJsonStr litellmAdditionalParams}}){{else}}{}{{/if}}
+    return LiteLLMModel(
+        client_args=client_args,
+        model_id="{{#if modelId}}{{modelId}}{{else}}bedrock/us.anthropic.claude-sonnet-4-5-20250514-v1:0{{/if}}",
+        params=params,
+    )
+{{/if}}

package/dist/assets/python/http/strands/base/model/mantle_compat.py

@@ -0,0 +1,21 @@
+from strands.models.openai_responses import OpenAIResponsesModel
+
+
+class MantleCompatResponsesModel(OpenAIResponsesModel):
+    """Workaround for Bedrock Mantle rejecting output_text in EasyInputMessage content arrays.
+
+    Mantle's Pydantic validation only accepts content as a plain string for assistant messages, while
+    real OpenAI accepts both formats. Flatten assistant content arrays to strings so multi-turn works.
+    Used for open-source OpenAI models (gpt-oss-*) on the /v1 Mantle path; proprietary models use the
+    plain OpenAIResponsesModel on /openai/v1.
+    """
+
+    @classmethod
+    def _format_request_messages(cls, messages):
+        formatted = super()._format_request_messages(messages)
+        for msg in formatted:
+            if msg.get("role") == "assistant" and isinstance(msg.get("content"), list):
+                msg["content"] = "".join(
+                    part.get("text", "") for part in msg["content"] if part.get("type") == "output_text"
+                )
+        return formatted

package/dist/assets/python/http/strands/base/pyproject.toml

@@ -16,6 +16,9 @@ dependencies = [
     {{#if (eq modelProvider "Gemini")}}"google-genai >= 1.0.0",
     {{/if}}"mcp >= 1.19.0",
     {{#if (eq modelProvider "OpenAI")}}"openai >= 1.0.0",
+    {{/if}}{{#if (eq modelProvider "LiteLLM")}}"litellm >= 1.0.0",
+    {{/if}}{{#if bedrockMantle}}"openai >= 1.0.0",
+    "aws-bedrock-token-generator >= 1.0.0",
     {{/if}}"strands-agents >= 1.15.0",
     {{#if (or hasBrowser hasCodeInterpreter)}}"strands-agents-tools >= 0.1.0",
     {{/if}}{{#if hasBrowser}}"nest-asyncio >= 1.5.0",

package/dist/assets/python/http/strands/capabilities/memory/session.py

@@ -28,7 +28,7 @@ def get_memory_session_manager(session_id: Optional[str], actor_id: str) -> Opti
         f"/episodes/{actor_id}/{session_id}": RetrievalConfig(top_k=5, relevance_score=0.5),
 {{/if}}
 {{#if (includes memoryProviders.[0].strategies "SUMMARIZATION")}}
-        f"/summaries/{actor_id}/{session_id}": RetrievalConfig(top_k=3, relevance_score=0.5),
+        f"/summaries/{actor_id}": RetrievalConfig(top_k=3, relevance_score=0.5),
 {{/if}}
     }
 {{/if}}

package/dist/assets/typescript/http/strands/base/main.ts

@@ -3,6 +3,9 @@ import { Agent, McpClient, tool, type ToolList } from '@strands-agents/sdk';
 import { z } from 'zod';
 import { loadModel } from './model/load.js';
 import { getStreamableHttpMcpClient } from './mcp_client/client.js';
+{{#if hasMemory}}
+import { getActorId, getOrCreateMemoryManager } from './memory/memory.js';
+{{/if}}
 
 // Define a collection of MCP clients (filter out anything that failed to initialize)
 const mcpClients: McpClient[] = [getStreamableHttpMcpClient()].filter(
@@ -31,34 +34,109 @@ const SYSTEM_PROMPT = `
 You are a helpful assistant. Use tools when appropriate.
 `;
 
-let cachedAgent: Agent | null = null;
+{{#if hasMemory}}
+const agentCache = new Map<string, Agent>();
 
-async function getOrCreateAgent(): Promise<Agent> {
-  if (!cachedAgent) {
-    const model = await loadModel();
-    cachedAgent = new Agent({
-      model,
-      systemPrompt: SYSTEM_PROMPT,
-      tools,
-    });
+async function getOrCreateAgent(sessionId: string, actorId: string): Promise<Agent> {
+  const key = `${actorId}:${sessionId}`;
+  let agent = agentCache.get(key);
+  if (agent) return agent;
+
+  const model = await loadModel();
+  agent = new Agent({
+    model,
+    systemPrompt: SYSTEM_PROMPT,
+    tools,
+    memoryManager: getOrCreateMemoryManager(sessionId, actorId) ?? undefined,
+  });
+  agentCache.set(key, agent);
+  return agent;
+}
+{{else}}
+const AGENT_CACHE_LIMIT = 128;
+
+// Reuses one Agent per sessionId so each session keeps its own in-process
+// conversation history (best-effort; resets on cold start). A Map preserves
+// insertion order, so it doubles as an LRU bounded to 128 sessions — a local
+// dev process serving many sessions cannot leak history between them or grow
+// without bound. On AgentCore Runtime each microVM serves a single session, so
+// this holds one entry. For durable history, attach memory.
+const agentCache = new Map<string, Agent>();
+
+async function getOrCreateAgent(sessionId: string): Promise<Agent> {
+  const existing = agentCache.get(sessionId);
+  if (existing) {
+    agentCache.delete(sessionId);
+    agentCache.set(sessionId, existing);
+    return existing;
   }
-  return cachedAgent;
+  if (agentCache.size >= AGENT_CACHE_LIMIT) {
+    const oldest = agentCache.keys().next().value;
+    if (oldest !== undefined) agentCache.delete(oldest);
+  }
+  const model = await loadModel();
+  const agent = new Agent({
+    model,
+    systemPrompt: SYSTEM_PROMPT,
+    tools,
+  });
+  agentCache.set(sessionId, agent);
+  return agent;
 }
+{{/if}}
 
 const app = new BedrockAgentCoreApp({
   invocationHandler: {
     async *process(payload: any, context: any) {
-      const agent = await getOrCreateAgent();
+      {{#if hasMemory}}
+      const sessionId = context?.sessionId ?? 'default-session';
+      const actorId = getActorId(payload, context);
+      const agent = await getOrCreateAgent(sessionId, actorId);
+      {{else}}
+      const sessionId = context?.sessionId ?? 'default-session';
+      const agent = await getOrCreateAgent(sessionId);
+      {{/if}}
 
-      for await (const event of agent.stream(payload.prompt ?? '')) {
-        if (
-          event.type === 'modelStreamUpdateEvent' &&
-          event.event?.type === 'modelContentBlockDeltaEvent' &&
-          event.event.delta?.type === 'textDelta'
-        ) {
-          yield { data: event.event.delta.text };
+      {{#if hasMemory}}
+      try {
+        for await (const event of agent.stream(payload.prompt ?? '')) {
+          if (
+            event.type === 'modelStreamUpdateEvent' &&
+            event.event?.type === 'modelContentBlockDeltaEvent' &&
+            event.event.delta?.type === 'textDelta'
+          ) {
+            yield { data: event.event.delta.text };
+          }
+        }
+      } finally {
+        // Drain in-flight createEvent calls before the runtime can reclaim
+        // the session microVM. flush() is the durability mechanism — without
+        // it, an idle reclamation can lose the tail of the conversation.
+        await agent.memoryManager?.flush();
+      }
+      {{else}}
+      // Snapshot history before streaming so a failed turn can be rolled back.
+      // Agent.stream() appends the user message before invoking the model; on a
+      // mid-stream error that user turn would otherwise linger in the cached
+      // agent, and the next turn for this session would send consecutive user
+      // messages (rejected by providers that require strict role alternation,
+      // e.g. Anthropic). Restoring on error keeps the session reusable.
+      const snapshot = agent.takeSnapshot({ include: ['messages'] });
+      try {
+        for await (const event of agent.stream(payload.prompt ?? '')) {
+          if (
+            event.type === 'modelStreamUpdateEvent' &&
+            event.event?.type === 'modelContentBlockDeltaEvent' &&
+            event.event.delta?.type === 'textDelta'
+          ) {
+            yield { data: event.event.delta.text };
+          }
         }
+      } catch (error) {
+        agent.loadSnapshot(snapshot);
+        throw error;
       }
+      {{/if}}
     },
   },
 });

package/dist/assets/typescript/http/strands/base/package.json

@@ -20,8 +20,9 @@
     "@google/genai": "^1.40.0",
     {{/if}}
     "@modelcontextprotocol/sdk": "^1.25.2",
-    "@strands-agents/sdk": "1.0.0-rc.4",
-    "bedrock-agentcore": "^0.2.4",
+    "@opentelemetry/api": "^1.9.0",
+    "@strands-agents/sdk": "^1.5.0",
+    "bedrock-agentcore": "^0.3.0",
     "tsx": "^4.19.0",
     "zod": "^4.4.3"
   },

package/dist/assets/typescript/http/strands/capabilities/memory/memory.ts

@@ -0,0 +1,52 @@
+import { randomUUID } from 'node:crypto';
+import { MemoryManager } from '@strands-agents/sdk';
+import { createAgentCoreMemoryStores } from 'bedrock-agentcore/experimental/memory/strands';
+
+const MEMORY_ID = process.env.{{memoryProviders.[0].envVarName}};
+
+const CUSTOM_ACTOR_ID_HEADER = 'x-amzn-bedrock-agentcore-runtime-custom-actor-id';
+
+export function getActorId(payload: any, context: any): string {
+  const raw =
+    context?.headers?.[CUSTOM_ACTOR_ID_HEADER] ||
+    payload?.userId ||
+    context?.sessionId;
+  return typeof raw === 'string' && raw.trim().length > 0 ? raw.trim() : randomUUID();
+}
+
+const memoryManagerCache = new Map<string, MemoryManager>();
+
+export function getOrCreateMemoryManager(sessionId: string, actorId: string): MemoryManager | null {
+  if (!MEMORY_ID) return null;
+
+  const key = `${actorId}:${sessionId}`;
+  let manager = memoryManagerCache.get(key);
+  if (manager) return manager;
+
+  const stores = createAgentCoreMemoryStores({
+    memoryId: MEMORY_ID,
+    actorId,
+    sessionId,
+    namespaces: [
+{{#if (includes memoryProviders.[0].strategies "SEMANTIC")}}
+      { namespace: '/users/{actorId}/facts' },
+{{/if}}
+{{#if (includes memoryProviders.[0].strategies "USER_PREFERENCE")}}
+      { namespace: '/users/{actorId}/preferences' },
+{{/if}}
+{{#if (includes memoryProviders.[0].strategies "EPISODIC")}}
+      { namespace: '/episodes/{actorId}/{sessionId}' },
+{{/if}}
+{{#if (includes memoryProviders.[0].strategies "SUMMARIZATION")}}
+      { namespace: '/summaries/{actorId}/{sessionId}' },
+{{/if}}
+    ],
+    // readMode defaults to 'per-namespace' (one retrieve call per namespace).
+    // Switch to 'subtree' to consolidate to a single hierarchical recall call.
+    extraction: true,
+  });
+
+  manager = new MemoryManager({ stores });
+  memoryManagerCache.set(key, manager);
+  return manager;
+}

package/dist/assets/typescript/http/vercelai/base/main.ts

@@ -1,22 +1,62 @@
 import { BedrockAgentCoreApp } from 'bedrock-agentcore/runtime';
-import { streamText } from 'ai';
+import { streamText, type ModelMessage } from 'ai';
 import { loadModel } from './model/load.js';
 
 const SYSTEM_PROMPT = `You are a helpful assistant.`;
 
+const HISTORY_LIMIT = 128;
+
+// Keeps one message history per sessionId so each session remembers its own
+// turns (best-effort; resets on cold start). A Map preserves insertion order,
+// so it doubles as an LRU bounded to 128 sessions — a local dev process serving
+// many sessions cannot leak history between them or grow without bound. On
+// AgentCore Runtime each microVM serves a single session, so this holds one
+// entry. For durable history, persist messages to an external store.
+const histories = new Map<string, ModelMessage[]>();
+
+function getHistory(sessionId: string): ModelMessage[] {
+  const existing = histories.get(sessionId);
+  if (existing) {
+    histories.delete(sessionId);
+    histories.set(sessionId, existing);
+    return existing;
+  }
+  if (histories.size >= HISTORY_LIMIT) {
+    const oldest = histories.keys().next().value;
+    if (oldest !== undefined) histories.delete(oldest);
+  }
+  const fresh: ModelMessage[] = [];
+  histories.set(sessionId, fresh);
+  return fresh;
+}
+
 const app = new BedrockAgentCoreApp({
   invocationHandler: {
     async *process(payload: any, context: any) {
+      const sessionId = context?.sessionId ?? 'default-session';
+      const history = getHistory(sessionId);
+      const userMessage: ModelMessage = { role: 'user', content: payload.prompt ?? '' };
+
       const model = await loadModel();
       const result = streamText({
         model,
         system: SYSTEM_PROMPT,
-        prompt: payload.prompt ?? '',
+        messages: [...history, userMessage],
       });
 
+      let assistant = '';
       for await (const chunk of result.textStream) {
+        assistant += chunk;
         yield { data: chunk };
       }
+
+      // Commit the exchange to history only after a non-empty reply. On a failed
+      // or empty stream the turn is dropped instead of leaving a dangling user
+      // (or empty assistant) message — consecutive same-role or empty-content
+      // messages would otherwise be rejected on the next turn for this session.
+      if (assistant.length > 0) {
+        history.push(userMessage, { role: 'assistant', content: assistant });
+      }
     },
   },
 });